Research map / Graph learning / GNN architectures
Neural Message Passing for Quantum Chemistry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed edge-network MPNN with set2set readout and explicit hydrogen nodes | MPNN · QM9, 13 targets | error ratio: state of the art on all 13 targets; chemical accuracy on 11 out of 13 targets | Helped |
| Ensemble the five models with lowest validation error | enn-s2s · QM9, 13 targets | error ratio: 0.68 -> 0.52 | Helped |
| Add spatial distance information to the molecular graph (ablation) | MPNN with set2set output · QM9, average across 13 targets | average error ratio: 2.57 -> 0.98 | Helped |
| Represent hydrogen atoms as explicit graph nodes (ablation) | MPNN with distance information · QM9, average across 13 targets | average error ratio: 0.98 -> 0.68 | Helped |
| Add virtual edges between unconnected node pairs (ablation) | GG-NN without spatial information · QM9, average across 13 targets | average error ratio: 3.47 -> 2.90 | Helped |
| Add a master node connected to every input node (ablation) | GG-NN without spatial information · QM9, average across 13 targets | average error ratio: 3.47 -> 2.62 | Helped |
| Use set2set readout (ablation) | GG-NN without spatial information · QM9, average across 13 targets | average error ratio: 3.47 -> 2.57 | Helped |
| Use eight towers instead of a vanilla GG-NN (ablation) | GG-NN with set2set and distance bins, individual training · QM9, average across 13 targets | average error ratio: 1.53 -> 1.37 | Helped |
| Use eight towers instead of a vanilla GG-NN (ablation) | GG-NN with set2set and distance bins, joint training · QM9, average across 13 targets | average error ratio: 1.92 -> 1.75 | Helped |
| Use an edge-network message function instead of pair messages (ablation) | MPNN in joint training · QM9, average across 13 targets | average error ratio: 3.98 -> 1.53 | Helped |
| Train a separate model per target rather than jointly predicting all 13 targets (ablation) | MPNN · QM9 | error ratio: up to 40% | Helped |
| Disable weight tying across message-passing time steps (ablation) | MPNN · QM9 | error ratio | Hurt |
Closest papers by meaning
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.87)
- $\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials (NeurIPS 2024 · similarity 0.86)
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 2024 · similarity 0.86)
- Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.86)
- Equivariant Matrix Function Neural Networks (ICLR 2024 · similarity 0.86)
- Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.